Detection of SARS-CoV-2 infection in the general population by three prevailing rapid antigen tests: cross-sectional diagnostic accuracy study

This large cross-sectional diagnostic test accuracy study was embedded within the Dutch public testing infrastructure. Public testing in the Netherlands is free-of-charge but only available for government-approved test indications. At the time of the study (12 April to 14 June 2021), these indications included having symptoms of suspected SARS-CoV-2 infection or having been identified as a close contact of a SARS-CoV-2 index case via traditional contact-tracing or the contact-tracing app regardless of symptomatology at the time of notification. Participants were recruited consecutively at three Dutch public health service COVID-19 test sites across the country, located in the West-Brabant region (Breda, using the BD-Veritor Ag-RDT), in the Rotterdam-Rijnmond region (Rotterdam The Hague Airport and Ahoy, using the SD-Biosensor Ag-RDT), and in the IJsselland region (Zwolle, using the PanBio Ag-RDT). Individuals were considered eligible if they were aged 16 years or older and willing and able to sign an informed consent in Dutch.

The Dutch COVID-19 vaccination programme started on 6 January 2021. At the time of the study, an estimated 20% (12 April) to 62% (14 June) of Dutch inhabitants aged 18 or older had received at least one vaccination, ranging over time from 4–13% for 18–25-year-olds to 85–91% for 81–90-year-olds [8]. The study was conducted before the SARS-CoV-2 Delta variant became the dominant variant in the Netherlands in June 2021 (the prevalence of the Delta variant was around 8.6% in the last week of inclusions) [9, 10].

Participants arrived at the test sites by car or bicycle (Breda) or on foot (Rotterdam and Zwolle). Test site staff verbally verified study eligibility. Eligible individuals were given a study flyer and a participant information letter to read, after which they could indicate to site staff if they wanted to participate. After signing the informed consent form, participants completed a short questionnaire on indication for testing, presence, type, and onset of symptoms; previous SARS-CoV-2 infections; and COVID-19 vaccination status (Additional file 1: Suppl. Material 1 [6, 11‐18]) while waiting for sampling.

A trained test site staff member took two swabs from each study participant: one for molecular reference testing and the other for the Ag-RDT. The molecular reference test was performed in a centralised laboratory in each region, whereas the Ag-RDT was performed at the test sites. In the first phase of the study, the Ag-RDT swabs were collected using the sampling method that was routinely used at the test site, i.e. deep NP for SD-Biosensor and PanBio, and superficial OP-N (about 2.5 cm deep) for BD-Veritor. In the second phase of the study, we evaluated the SD-Biosensor and PanBio tests using superficial OP-N sampling. Ag-RDTs were conducted and interpreted in accordance with the manufacturer’s instructions; results of the BD-Veritor Ag-RDT were determined visually instead of using a Veritor Plus Analyzer [11].

While molecular testing was used as the reference standard in all three centralised laboratories, the sampling and molecular testing details varied slightly (Additional file 1: Suppl. Material 2). In Breda and Zwolle, OP-N sampling was combined with RT-PCR or transcription-mediated amplification (TMA) testing, respectively. Samples that tested positive by TMA in Zwolle were subsequently tested by RT-PCR to generate a Ct value. The Rotterdam site used combined oropharyngeal and nasopharyngeal (OP-NP) sampling combined with RT-PCR. The platforms used were Roche cobas 6800/8800 (Rotterdam and Breda, respectively) and ABI-7500 (Zwolle) for RT-PCR and the Hologic Panther system (Aptima SARS-CoV2 assay) for TMA (Additional file 1: Suppl. Material 2).

All staff assessing test results were blinded to the results of the other test. In the first phase of the study, the Ag-RDTs were conducted in accordance with routine test site procedures; participants were therefore informed about the Ag-RDT result but not the subsequent molecular test result. In the second phase of the study, the Ag-RDTs were not conducted according to routine practice and participants were therefore informed about the molecular test result.

In discordant cases (Ag-RDT-negative and RT-PCR-positive cases), whole genome sequencing (WGS) of the primary clinical sample was performed when the viral load was above a cut-off of ≥5.2 log10 SARS-CoV-2 E-gene copies/mL. This is the viral load above which 95% of people with a positive molecular test had a positive virus culture in a recent study by our group [6] (Additional file 1: Suppl. Material 2).

The primary outcomes were the diagnostic accuracies (sensitivity, specificity, positive and negative predictive values with corresponding 95% confidence intervals [CI]) of all three Ag-RDTs and sampling technique combinations, with molecular testing as the reference standard. As the number of individuals without molecular test or Ag-RDT results was very low (n=55 (0.7%); Fig. 1), we performed a complete case analysis.

Secondary outcomes were diagnostic accuracies above the viral load cut-off of ≥5.2 log10 SARS-CoV-2 E-gene copies/mL [6]. Additional secondary outcomes were diagnostic accuracies stratified by the presence of symptoms at time of sampling (yes or no), COVID-19 vaccination status (vaccinated with at least one dose yes or no), having had a prior SARS-CoV-2 infection (yes or no), sex (female or male), age (≥16 to ≤40 or >40 to ≤65 or >65), and testing indication (symptoms and/or close contact without symptoms). In an exploratory analysis, we performed WGS to assess whether false-negative Ag-RDT results could be linked to SARS-CoV-2 variants or specific mutations in the SARS-CoV-2 N-gene (Additional file 1: Suppl. material 2).

Finally, we used the SARS-CoV-2 test result database of the public health service test sites to identify any missed infections using pseudonymised linkage. Specifically, we determined whether participants who received a negative test result had tested positive in the subsequent 14 days by either molecular test or Ag-RDT, and analysed the interval between the initial and follow-up test. Follow-up results of participants were stratified by the results of the initial Ag-RDT and molecular reference tests and by the presence of symptoms at the time of sampling.

Previous diagnostic accuracy studies of Ag-RDTs in people with COVID-19-like symptoms found sensitivities of around 80–85% [3‐5, 11, 19]. We based our sample size calculation on an expected sensitivity of 80% for each Ag-RDT, with a margin of error of 7%, type I error of 5%, and power of 90%. We required approximately 145 positive reference tests for each Ag-RDT-molecular reference test comparison and per Ag-RDT sampling technique (routinely used versus less invasive). We expected a negligible non-response rate based on previous studies. We anticipated a SARS-CoV-2 prevalence (based on molecular testing) of 10% and closely monitored molecular test positivity rates over time in order to prolong recruitment as needed.

SARS-CoV-2 prevalence (by molecular reference test) was 13.0% (188/1441) in the BD-Veritor group, 12.2% (215/1769) in the SD-Biosensor group, and 8.4% (173/2056) in the PanBio group. Overall sensitivities were 68.6% [61.5–75.2%] for BD-Veritor, 74.4% [68.0–80.1%] for SD-Biosensor, and 68.8% [61.3–75.6%] for PanBio (Table 2, Fig. 2).

Among those with a positive molecular test result, the percentage of participants with a viral load above the cut-off was 77.7% (146/188) in the BD-Veritor group, 85.1% (183/215) in the SD-Biosensor group, and 70.0% (121/173) in the PanBio group. Using this viral load cut-off, the sensitivities were 85.6% [78.9–90.9%] for BD-Veritor, 87.9% [82.3–92.3%] for SD-Biosensor, and 89.3% [82.3–94.2%] for PanBio.

Sensitivities ranged from 72.2 to 83.4% in individuals who were symptomatic at the time of sampling and from 54.0 to 55.9% in those who were asymptomatic (Table 2, Fig. 2). We found no evidence of a differential impact on diagnostic accuracy by COVID-19 vaccination status, sex, and age (Fig. 2, Additional file 1: Suppl. Table S3).

Specificities were >99%, and positive and negative predictive values were >95%, for all three Ag-RDT in most analyses (Table 2 and Additional file 1: Suppl. Table S3).

The SARS-CoV-2 prevalence (by molecular reference test) was 9.7% (164/1689) and the sensitivity was 75.0% [67.7–81.4%] (Table 2, Fig. 2).

Among those with a positive molecular test result, the percentage of participants with a viral load above the cut-off was 86.0% (141/164). Using this viral load cut-off, the sensitivity was 83.7% [76.5–89.4%].

Sensitivities were 78.9% in symptomatic individuals and 63.0% in those who were asymptomatic at the time of sampling (Table 2, Fig. 2). We found no evidence of a differential impact on diagnostic accuracy by COVID-19 vaccination status, sex, and age (Additional file 1: Suppl. Table S3).

Specificities were >99%, and positive and negative predictive values were >95%, in most analyses (Table 2 and Additional file 1: Suppl. Table S3).

All but six of all study participants could be linked with the national test results database of the public health services for the follow-up analyses.

In the first phase of the study, participants received the result of the Ag-RDT test and not the result of the molecular reference test. This Ag-RDT result was false-negative in 3.5% of participants (Table 3). Of the participants who tested Ag-RDT-negative in the study, 18.3% had another SARS-CoV-2 test done at a public health service testing site within 14 days, and 4.4% tested positive on that repeat test (Table 3). These percentages were much higher for those with a false-negative than a true-negative Ag-RDT result during the study: 61.3% versus 16.8% (χ² =215, p<.001) for having a repeat test and 55.4% versus 2.6% (χ² =1076, p<.001) for that repeat test being positive (Table 3). Furthermore, the interval between the Ag-RDT-negative result during the study and the subsequent positive test results within the 14-day follow-up period was shorter for participants with false-negative results during the study (median 3 days, interquartile range (IQR) 3 days) than for those with true-negative results (median 5 days, IQR 3 days; U=2810, p<.001). Finally, being asymptomatic at the time of initial testing was associated with a higher likelihood of testing positive during the 14-day follow-up period for both false-negative and true-negative individuals (Table 3).

In the second phase of the study, participants received the molecular reference test result and not the Ag-RDT test result. Of the participants who tested reference test negative in the study, 11.5% had another SARS-CoV-2 test done at a public health service testing site within 14 days, and 1.1% tested positive on that repeat test (Table 3). The median interval between the initial negative reference test and a positive follow-up test was 6 days (IQR 5 days). Participants who were asymptomatic at the time of initial testing were more likely to test positive during the 14 day follow-up period than those who were symptomatic, but the difference was not statistically significant (Table 3).

The overall unstratified sensitivities of the Ag-RDTs in our study were substantially lower than those reported in Ag-RDT evaluations performed earlier in the pandemic in the Netherlands [11, 19]. We hypothesise that the reason for this is because only symptomatic individuals could access SARS-CoV-2 testing in the Netherlands until 1 December 2020. The sensitivities in our study for individuals who were symptomatic at the time of sampling were indeed similar to those in the earlier evaluation studies. The sensitivities in our study for individuals who were asymptomatic at the time of sampling are in line with those that we found in our recent study in asymptomatic and presymptomatic close contacts [6]. The positivity percentages during the 14-day follow-up period of people with a negative initial molecular test in our study was 1.1%, which was slightly lower than the 1.7% found in our previous study among close contacts. This can likely be attributed to the fact that close contacts are at higher risk of testing SARS-CoV-2-positive than those testing for other reasons [6].

To our knowledge, a direct comparison of the diagnostic accuracy of any Ag-RDT combined with superficial OP-N versus deep NP sampling has not been conducted. The superficial OP-N sampling technique is currently routinely used in the Netherlands for the BD-Veritor Ag-RDT. The sensitivities that we observed for this combination were similar to those of the other two Ag-RDTs combined with deep NP sampling, as well as the SD-Biosensor Ag-RDT combined with OP-N sampling. The use of more convenient sampling methods holds potential for reducing the threshold for SARS-CoV-2 testing.

Strengths include the protocolised nature of the study, the large sample size covering multiple test sites nationwide, the high data completeness, collection of samples for the index and reference tests at the same time, the implementation of index and reference tests by trained staff who were blinded to the result of the other test, and the availability of follow-up information for participants who received negative test results.

Our study also has some limitations. First, our study does not provide a direct comparison of the diagnostic accuracy of the various Ag-RDTs or sampling techniques. However, we did use similar eligibility criteria and study protocols across the three study sites, and we included participants at the three study sites during the same time period. Second, the reference standards that we used were molecular tests, but platforms and test kits used differed among the three centralised laboratories (Rotterdam: OP-NP sampling with RT-PCR (Roche cobas 6800), Breda: OP-N sampling with RT-PCR (Roche cobas 8800), Zwolle: OP-N sampling with TMA (Hologic Panther system) with TMA-positive samples tested by RT-PCR (ABI-7500)). However, the diagnostic accuracies of all molecular tests used are similarly high [20, 21], and we therefore believe that this has not influenced our findings significantly. In addition, Ct values determined by the different platforms were comparable (Additional file 1: Suppl. Material 2). Third, we were unable to meet the predefined sample size in the PanBio/OP-N sampling group. These results are therefore not sufficiently robust and should be interpreted with great caution. Fourth, while the Omicron variant is the dominant SARS-CoV-2 variant in the Netherlands at time of writing, this variant was not present during our study period and the estimated prevalence of Delta was 8.6% during the last week of inclusions. We checked whether false-negative Ag-RDT results could be linked to specific virus lineages by WGS and we did not find a signal confirming this hypothesis. Fifth, we applied a viral load cut-off above which 95% of RT-PCR test positive individuals in our previous study was virus culture positive [6]. The previous study population, however, consisted of almost completely unvaccinated participants, whereas the proportion of vaccinated individuals in the present study reached 20% at the end of the study. Whether this would have impacted the applied viral load cut-off is unknown. Also, the applied viral load cut-off should be interpreted with some caution. Albeit evidence is accumulating that both infectivity in culture and viral load, and viral load and secondary attack rate, are correlated [22‐24], the exact viral load cut-off below which no transmissions take place is still unknown. Missed infections below the applied viral load cut-off can therefore not be discounted as of limited relevance, especially because viral loads may alter over time during the infection. Sixth, our sample size calculation was based on the primary analysis and the diagnostic accuracy parameters are therefore less precise for the secondary stratified analyses. For example, a differential impact of COVID-19 vaccination status on the diagnostic accuracy of Ag-RDTs might be anticipated given that vaccinated individuals have shorter periods of high viral loads and lower virus viability by Ct values than unvaccinated individuals, resulting in lower transmissibility [24‐28]. Further studies on the potential impact of COVID-19 vaccination on Ag-RDT diagnostic accuracies are warranted. Finally, we did not actively follow-up participants who had received a negative test result but collected follow-up information from the public health services test result database through pseudonymised linkage. Active follow-up, including repeat testing in all study participants, would have reduced the uncertainty around false-negative Ag-RDT results completely, as was also recommended in a recent guidance paper [29]. Unfortunately, we could not implement this because of ethical and logistic constraints, as our study was embedded in busy public health service test sites. Also, we cannot be certain that all positive tests within the maximum incubation period after a negative initial test represent false-negative tests; they could also have resulted from a new SARS-CoV-2 exposure after the initial test. However, this is true for individuals with a negative Ag-RDT and for those with a negative molecular test at inclusion, and the difference in positivity percentages during follow-up was strikingly large.

As the COVID-19 pandemic will gradually move into a different phase with less emphasis on finding all infections, public health policies may want to rely more on Ag-RDTs instead of molecular testing because of simplified logistics and reduced delays. The more frequent use of Ag-RDTs instead of molecular testing will inevitably lead to an increase in the number of missed infections, especially when used in asymptomatic individuals, e.g. for travelling and/or access to events. We observed that about 60% of persons with false-negative Ag-RDT results returned for testing, and 55% tested positive, within 14 days after the initial negative test with a median delay of 3 days. About half of them did not have symptoms at the time of initial sampling, and those participants were more likely to have a positive follow-up test (symptomatic vs asymptomatic at time of initial sampling: 46% vs 63%). This suggests that a considerable portion of individuals adhered to the advice to return for testing if symptoms develop or worsen after a negative test and also illustrates the importance of emphasizing the need for re-testing to individuals with a negative Ag-RDT should such events occur. However, as over half of all infections are estimated to occur before the onset of symptoms, these missed early infections do pose a relevant risk for further transmission [30, 31]. Furthermore, the 45% of individuals with a false-negative Ag-RDT who are never diagnosed pose an even greater risk of onward transmission. The extent to which the advantages of Ag-RDTs outweigh their lower sensitivities depends on several aspects, including the potential consequences of missed infections.

Furthermore, as the prevalence of SARS-CoV-2 declines, the positive predictive value of Ag-RDT results will decrease, meaning a larger proportion of positive test results will be false positive [3]. In such circumstances, the risk of false-positive results with Ag-RDTs could be mitigated by confirmatory molecular testing.